Search for optimal steering geometry by systematically exploring combinations of layers, extraction strategies, and benchmarks. This command helps identify the best configuration for creating effective steering vectors.
python -m wisent geometry-search --model MODEL [OPTIONS]
python -m wisent geometry-search \ --model meta-llama/Llama-3.1-8B-Instruct \ --output ./geometry_results.json
python -m wisent geometry-search \ --model meta-llama/Llama-3.1-8B-Instruct \ --benchmarks arc_easy,hellaswag,truthfulqa_mc1 \ --pairs-per-benchmark 100 \ --output ./benchmark_geometry.json
python -m wisent geometry-search \ --model meta-llama/Llama-3.1-8B-Instruct \ --strategies chat_last,chat_mean,chat_max_norm \ --max-layer-combo-size 2 \ --output ./strategy_geometry.json
python -m wisent geometry-search \ --model meta-llama/Llama-3.1-8B-Instruct \ --max-layer-combo-size 3 \ --pairs-per-benchmark 50 \ --cache-dir ./geometry_cache/ \ --seed 42 \ --output ./full_geometry.json
python -m wisent geometry-search \ --model meta-llama/Llama-3.1-8B-Instruct \ --benchmarks ./my_benchmarks.txt \ --output ./custom_geometry.json
| Argument | Default | Description |
|---|---|---|
| --model | required | Model name or path |
| --output | geometry_results.json | Output path for results JSON |
| --pairs-per-benchmark | 50 | Number of pairs to sample per benchmark |
| --max-layer-combo-size | 3 | Maximum layers in combination (1=individual, 2=pairs, 3=triplets) |
| --strategies | all (7) | Comma-separated list of extraction strategies |
| --benchmarks | all | Comma-separated benchmarks or path to .txt file |
| --cache-dir | /tmp/wisent_geometry_cache | Directory for activation cache |
| --seed | 42 | Random seed for reproducibility |
| --device | auto | Device for model (auto/cuda/mps/cpu) |
The search produces a JSON file containing:
Stay in the loop. Never miss out.
Subscribe to our newsletter and unlock Wisent insights.